Small is good. DTU and KU have created the sought-after and completely sold-out workshop, which goes against the stream and provides space for real knowledge sharing in the AI environment.
90 years ago, Copenhagen was the centre of Niels Bohr's research in quantum physics, and people traveled a long way to participate. This is somewhat the same ambitions that researchers from DTU Compute and the Department of Computer Science at the University of Copenhagen are pursuing within the special research area of AI called generative models. That is, AI tools analyze the underlying mechanisms of artificial intelligence to make it more possible to understand why AI gives a certain result.
85 AI researchers have just gathered for two days to exchange knowledge about AI models in the very same buildings of the Carlsberg Academy in Valby, where Niels Bohr lived for more than 30 years until his death in 1962.
The very special thing about the workshop was that it only had room for 85 participants.
"You always talk about research being 10-15 years away. That is not the case with Machine Learning. Look half a year into the future, and then your research is almost outdated. Development is going extremely fast. If you don't look at others' research and chat with people for a few years, then you are out of the question as a researcher. That is why we have to talk to people in an informal setting. And traditionally, conferences have had that role. But conferences no longer serve that purpose," says Professor at DTU Compute Søren Hauberg, one of the organizers.
"Today, Machine Learning conferences have many thousands of participants, and workshops maybe 500 participants. That kills easily all discussions. Because when you are faced with such a large crowd, you are not so willing to discuss weaknesses in your models. You almost go into defensive mode and sell your already published article. So conferences are no longer a place where you challenge each other and try to get better.”
"You always talk about research being 10-15 years away. That is not the case with Machine Learning. Look half a year into the future, and then your research is almost outdated. Development is going extremely fast. If you don't look at others' research and chat with people for a few years, then you are out of the question as a researcher. That is why we have to talk to people in an informal setting."
Søren Hauberg, Professor at DTU Compute
So the Copenhagen meeting 'genu.ai' is thought of completely differently. A little more than 50 percent of the time is set aside for questions and discussion, explains Søren's colleague, Associate Professor Jes Frellsen:
"Originally, researchers brought a scientific report and discussed it with the audience and then published it after the conference. We want to go back to that stage, where people present their work in progress and return home to finish the article with better models."
Missed European network
The AI Copenhagen meeting is intended as a European meeting because, in addition to a few individual presenters from the USA, the participants come from Europe because the universities want to strengthen the research environment in Europe.
One of the presenters was Carl Henrik Ek, Associate Professor at the Computer Laboratory at the University of Cambridge and Associate Professor in Machine Learning at KTH - the Royal Institute of Technology in Stockholm. He came straight from a summer school in England, where he had given a lecture to young researchers in the morning. He is very enthusiastic about the entire setup for the AI Copenhagen meeting, where the exclusive circle of participants also has room for very young researchers, so that the research environment does not close in on itself.
"I have missed a network and being able to gather in Europe. So far we have been very oriented towards what was happening in the USA. Perhaps it was necessary because the research environment was very small in the past, but that has changed. Today, we have enough researchers in the research field, e.g. in the Nordic countries. And the payoff here is much better than at large conferences because we dare to say something that we don't know for sure is true, but that others know about. So together we find new knowledge. And the research needs that," says Carl Henrik Ek.
Carl Henrik Ek at genu.ai.
Successful 'maternity project'
This year it was the third time the workshop took place. The first time was in 2019. At that time, several of the Danish organizers had small children, and it was quite impossible to travel to conferences.
Instead, they came up with inviting people to Copenhagen. And it has been a great success. Although it was never advertised, there was a waiting list.
"We have received inquiries from people we don't know, who wrote that they had heard this was the 'meeting' - and they would like to join us. We have had to turn away many because we, unfortunately, had no more places. It's super sad, but also super cool that our initiative that started as a maternity project has de facto turned into something that people are queuing up for," says Søren Hauberg.
Genu.ai has received financial support from Center for Basic Machine Learning Research in Life Science (MLLS), Pioneer Centre for AI, Independent Research Fund Denmark’s (DFF), European Research Council (ERC), and Carlsberg Foundation.
Poster session at genu.ai.
- The theme of Genu.ai is generative models and uncertainty quantification.
- The research area is a combination of deep learning – networks that are said to be inspired by the brain – and probability theory.
- The generative models are hot stuff these days because they can create synthetic images. You can ask the AI to create an image by describing what should be in the image, and the model then creates a new image using images from the web. The models can also create voices and are the basic tool in deep fake.
- The problem with many of these models is that they have difficulties expressing how much uncertainty there is. It can be a serious problem, e.g. if you want to use models to find patterns in the development of cancer cells. In those scenarios, it is important to be able to describe the uncertainty. Models are able to describe uncertainty, but the researchers would like the model to tell when one can understand one and the other from data.
- The Genu.ai conference is about the intersection, so how to get generative models to describe the process by which data could have been created.